fix(ai): reject WebSocket transport for OpenAI Chat #40655

Problem

The OpenAI Chat provider entrypoint shares Settings with the Responses entrypoint, so it accepts transport: "websocket". However, chatModel() ignores settings.transport and always selects configure(config(settings)).chat(modelID), whose /v1/chat/completions route uses HTTP/SSE.

This makes a WebSocket selection on @opencode-ai/ai/providers/openai/chat silently behave like the default transport. OpenAI documents persistent WebSocket mode for the Responses API, not Chat Completions:

https://developers.openai.com/api/docs/guides/websocket-mode

Relevant code

• packages/ai/src/providers/openai.ts: shared Settings.transport and chatModel()
• packages/ai/src/providers/openai/chat.ts: exports chatModel as model
• packages/ai/src/protocols/openai-chat.ts: HTTP/SSE Chat route

Expected behavior

Make unsupported transport selection impossible or explicit for the Chat entrypoint. Prefer a Chat-specific settings surface that does not expose transport; if runtime input can still contain it, reject transport: "websocket" with a clear error instead of silently ignoring it.

Acceptance criteria

• The OpenAI Chat package entrypoint no longer silently accepts WebSocket transport.
• Type-level coverage prevents WebSocket selection at normal TypeScript call sites.
• Runtime coverage verifies a clear failure for untyped/serialized unsupported input, if that input reaches this boundary.
• OpenAI Responses keeps its existing "http" | "websocket" setting.